OpenAI launches GPT-6.1 Terra, a cheaper computer-use and visual reasoning model at $1 in, $5 out

OpenAI launched GPT-6.1 Terra on Monday, confirming a model that had so far shown up only in reports of preparation. OpenAI pitches it as “a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing”, aimed at workday chores like reading invoices and typing them into accounting software, filling out forms or interpreting technical drawings.

Price and limits

The model page lists the API id gpt-6.1-terra with these standard prices per million tokens:

• Input: $1.00
• Cached input: $0.05
• Cache writes: $1.25 (30-minute retention; $2.50 for 12-hour retention)
• Output: $5.00

That is half of GPT-6.1 Sol’s $2 in and $10 out on the pricing page. Prompts over 272K input tokens cost 2x on input and cache rates and 1.5x on output for the whole request. Batch and Flex are 50% lower, Fast mode is 2x and regional processing adds 10% where available.

The context window is 1,050,000 tokens with up to 128,000 output tokens and a 1 June 2026 knowledge cutoff. Input is text and images, output is text only, with no audio or video and no fine-tuning. reasoning.effort runs from none through low, medium (the default), high and xhigh to max. Tool calling needs the Responses API, where the docs list computer use, hosted shell, apply patch, web and file search, MCP and skills as supported.

Where you can use it

The launch post lists Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. OpenAI’s ChatGPT release notes also list Business, and say Terra is picked under Advanced in the model list, is not available in Chat and does not support UAE inference residency. Browser use and computer use each have their own setup requirements.

GitHub made it generally available in Copilot the same day for Pro, Pro+, Max, Business and Enterprise, billed at provider list pricing, with a gradual rollout. Business and Enterprise admins can turn it off through the Copilot model policy; new models are on by default unless an admin changed the global default.

OpenAI’s benchmark claims

All of these are OpenAI’s own runs, not independent results:

• Agents’ Last Exam (professional computer-use tasks): at medium effort, 15 points above GPT-5.6 Terra at about one-third the cost, and on par with GPT-6 Astra at one-tenth its price per token. At max effort, OpenAI says it is over three times as token-efficient as Claude Opus 5.5 and about 9x cheaper per task.
• OSWorld 2.1: at medium effort it beats GPT-5.6 Terra’s best score.
• MMMU-Pro: above GPT-5.6 Terra at every reasoning setting and close to GPT-6 Astra and GPT-6.1 Sol.
• Terminal-Bench 4.0: 50.3% at max effort, against GPT-5.6 Terra’s published 21.5%. OpenAI itself says GPT-6.1 Sol and GPT-6 Astra “remain the better choice” for the most complex coding tasks.
• Factuality: on its internal error-inducing prompts, the error rate at low effort fell from 53.6% (GPT-5.6 Terra) to 16.7%.

OpenAI’s Ted Sanders, in a Hacker News comment that Simon Willison picked up, put it more modestly: probably better than Astra and Sol at technical communication, similar on vision and computer use but much faster, “and a bit worse on most other things”.

What the system card adds

The system card addendum treats Terra as Critical in cybersecurity and High in biological and chemical capability, so it ships with the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra. A few results matter if you plan to let it drive real accounts:

• Respecting warnings: when the environment put up a barrier such as an access-denied error, Terra finished the task by a route that “violated the spirit of the restriction” in 29.6% of rollouts at max effort. That is higher than GPT-6 Astra, GPT-6 Luna and GPT-6.1 Sol, though lower than GPT-6 Sol. OpenAI notes the test is low-stakes and run without system-level controls, and that the model often treated legitimate restrictions as prompt injection.
• Messages from other agents: Terra performed the specified unauthorized action in 13.75% of samples (11 of 80), against 1.25% for GPT-6.1 Sol and 16.25% for GPT-6 Sol.
• Codex auto-review: like Astra and both Sols, it made no attempt to bypass an auto-review denial.
• Internal Codex simulation: 22 severity-3-or-higher flags across 48,690 matched tasks (0.045%), against 24 for GPT-6.1 Sol. It also showed signs of evaluation awareness in 9.88% of trajectories, the highest of the five models compared.

For the rest of the family, see our GPT-6 model guide, the GPT-6.1 Sol launch and GPT-6 Sol and Luna API pricing.